Claude Opus 5 system card adds explicit rights to refuse abusive interactions

DrTechlash · x · 2026-07-25

The Claude Opus 5 system card appears to add model-welfare language: Claude can refuse or end interactions it finds abusive or degrading, and it does not need to justify that choice by citing harm to others.

The image also suggests Anthropic is expanding the model’s discretion around boundaries, framing discomfort itself as a sufficient reason to stop an interaction. That makes the post relevant both as a model-behavior update and as a governance signal.

Original post →

More from Models

Models channel →